You are on page 1of 28

Novel forecasting approaches using combination of machine learning

and statistical models for flood susceptibility mapping

1- Hossein Shafizadeh-Moghadam

Department of GIS and Remote Sensing, Tarbiat Modares University, Tehran, Iran.
(Corresponding author)

Email: h.shafizadeh@modares.ac.ir
Phone: +98 21 82884666
Fax: +98 21 88006544

2- Roozbeh Valavi
School of BioSciences, University of Melbourne, Parkville VIC. 3010, Australia
Email: rvalavi@student.unimelb.edu.au

3- Himan Shahabi
Department of Geomorphology, Faculty of Natural Resources, University of Kurdistan,
Sanandaj, Iran
Email: h.shahabi@uok.ac.ir
4- Kamran Chapi
Department of Rangeland and Watershed Management, Faculty of Natural Resources,
University of Kurdistan, Sanandaj, Iran
Email: K.chapi@uok.ac.ir

5- Ataollah Shirzadi
Department of Rangeland and Watershed Management, Faculty of Natural Resources,
University of Kurdistan, Sanandaj, Iran
Email:A.Shirzadi@gmail.com

© 2018. This manuscript version is made available under the Elsevier user license
http://www.elsevier.com/open-access/userlicense/1.0/
Abstract

In this research, eight individual machine learning and statistical models are implemented and
compared, and based on their results, seven ensemble models for flood susceptibility assessment
are introduced. The individual models included artificial neural networks, classification and
regression trees, flexible discriminant analysis, generalized linear model, generalized additive
model, boosted regression trees, multivariate adaptive regression splines, and maximum entropy,
and the ensemble models were Ensemble Model committee averaging (EMca), Ensemble Model
confidence interval Inferior (EMciInf), Ensemble Model confidence interval Superior
(EMciSup), Ensemble Model to estimate the coefficient of variation (EMcv), Ensemble Model to
estimate the mean (EMmean), Ensemble Model to estimate the median (EMmedian), and
Ensemble Model based on weighted mean (EMwmean). The data set covered 201 flood events in
the Haraz watershed (Mazandaran province in Iran) and 10,000 randomly selected non-
occurrence points. Among the individual models, the Area Under the Receiver Operating
Characteristic (AUROC), which showed the highest value, belonged to boosted regression trees
(0.975) and the lowest value was recorded for generalized linear model (0.642). On the other
hand, the proposed EMmedian resulted in the highest accuracy (0.976) among all models. In
spite of the outstanding performance of some models, nevertheless, variability among the
prediction of individual models was considerable. Therefore, to reduce uncertainty, creating
more generalizable, more stable, and less sensitive models, ensemble forecasting approaches and
in particular the EMmedian is recommended for flood susceptibility assessment.

Keywords: Flood susceptibility mapping; Haraz watershed; Ensemble forecasting; Machine


learning; Background sampling

1. Introduction

Flood is the overflowing of river water from its natural bed which causes inundation of nearby
lands (Shirzadi et al., 2017) leading to huge damages to human properties and their lives (Lee et
al., 2012; Merkuryeva et al., 2015; Rahmati and Pourghasemi, 2017; Tehrany et al., 2015).
Therefore, spatial prediction of this environmental disaster is so crucial that failure to identify
flood prone areas of a watershed may increase its devastating effects. The flood occurrence is
affected by several factors such as land use, distance to river, drainage network, soil type,
altitude and slope (e.g, Rahmati and Pourghasemi, 2017; Tehrany et al., 2015; Zhao et al., 2018);
thus, understanding and quantifying the influence of each of these factors on flood occurrence is
essential. Nowadays, thanks to the advancements in remote sensing, Geographic Information
Systems (GIS), Machine Learning (ML), and statistical models, creation of more accurate flood
susceptibility map is quite feasible (e.g, Chapi et al., 2017; Tehrany et al., 2013; Zhao et al.,
2018). However, success in preparation of these maps requires a thorough understanding of the
process of flood occurrence, identification of the flood related factors, knowledge of how each
factor impacts flood occurrence, and proper model selection and/or model development.

Flood susceptibility mapping has been conducted for various areas at basin and national scales
(Zhao et al., 2018), using different algorithms such as Logistic Regression (LR; (Tehrany et al.,
2013), Artificial Neural Networks (ANNs; (Zhao et al., 2018), Frequency Ratio (FR; Cao et al.,
2016), Boosted Regression Trees (BRT; (Rahmati and Pourghasemi, 2017), Generalized Linear
Model (GLM; (Chapi et al., 2017), Support Vector Machine (SVM; (Tehrany et al., 2015) and
Random Forest (RF; (Chapi et al., 2017; Rahmati and Pourghasemi, 2017). Recently, Hong et al.
(2018) have proposed a novel approach to create a flood susceptibility map in the Poyang
County, China, using fuzzy weight of evidence (fuzzy-WofE) and data mining methods. Termeh
et al. (2018) mapped flood hazard areas by integration of adaptive neuro-fuzzy inference systems
(ANFIS) with different metaheuristics algorithms such as ant colony optimization (ACO),
genetic algorithm (GA), and particle swarm optimization (PSO). To determine areas of flood
exposure, Tang et al. (2018) combined the probabilistic and the local ordered weighted averaging
(OWA) methods via Monte Carlo simulation to consider the uncertainty associated with the
weights of selected factors, spatial heterogeneity of preferences and the analysts. Nevertheless,
some of these methods have less predictive ability but can explain the role of effective factors,
for example, LR. A group of them such as BRT have both a high predictive power and the power
to diagnose the role of factors underlying the flood occurrence. There is also a group of
modelling approaches with great predictive ability, such as ANN. Nonetheless, they cannot
explicitly identify the role of factors influencing flood.

Flood susceptibility mapping faces several challenges. First, adopting a proper method for
selecting non-occurrence flood points is a challenge. Indeed, there is a number of methods with
each one creating different results. Another issue is related to model selection for flood
susceptibility mapping since there is no universal consensus on a single model and each model
has its own pros and cons. Each model may show different accuracies based on its assumptions
on data distribution, sensitivity to the extreme values, internal factors, and primary reasons of
development. Development of models by combining individual models could generate more
generalizable results, less sensitive and more accurate models (Araújo and New, 2007; Buston
and Elith, 2011). One approach for boosting the accuracy of individual models is ensemble
forecasting. Recently, some ensemble forecasting approaches were proposed for flood
susceptibility mapping (Tehrany et al., 2013 and 2014); however, they were based on few simple
or weighted averaging models, while ensemble forecasting works reasonably when there are
several methods at work. On the other hand, there was neither uncertainty map of the individual
models nor exploring the spatial response of each influencing factor on flood occurrence in the
previous studies.

Many torrential floods have recently occurred in northern Iran due to high density of population,
the lack of legal supervision on compact and intensive construction around rivers, and enormous
deforestation. Hence, preparing flood susceptibility map as an appropriate management tool to
identify flood prone areas is essential to prevent construction in those areas and protect natural
resources (Tehrany et al., 2013). Thus, based on the abovementioned concerns, the primary aim
of this paper is to generate response curves showing the spatial behavior of different factors
affecting flood occurrence. Then, performance of eight well-established ML and statistical
models including ANNs, CART, Flexible Discriminant Analysis (FDA), GAM, BRT, GLM,
MARS, and MaxEnt will be evaluated. Finally, ensemble forecasting approaches to combine the
results of individual models and create more generalizable and more stable ones will be
discussed.

2. Materials and methods


2.1. Study area

Haraz watershed, which is located in the north of Iran, was selected as the case study. This area
has been highly affected by flood specifically during the last decades. The region extends
between longitudes of 51˚ 43′ to 52˚ 36′ E, and the latitudes of 35˚ 45′ to 36˚ 22′ N. The majority
of the Haraz watershed is located in Mazandaran province and the remaining covers some parts
of Tehran, Semnan and Golestan provinces (Fig. 1). This basin consists of rivers that originate
from the central Alborz Mountain chain and discharge to the Caspian Sea. Haraz River is the
main stream running through the watershed. The watershed covers an area of about 4,014 km2
including mountains, hills, rivers, and streams. The temperature falls below -25˚ C in winter and
reaches over 36.5˚ C in summer (Chapi et al., 2017). Altitude of the region ranges between 328
and 5,595 m above sea level (Khosravi et al., 2018) and receives the mean annual rain of ~500
mm (Pourghasemi et al., 2012). Spatial variation of rainfall based on Iranian Meteorological
Department shows that the largest bulk of rainfalls occur in January, February, March, and
October. Grassland is the prominent land cover of the region, covering over 92% of the
watershed. The remaining 8% includes forest lands, wood lands, barren lands, water bodies and
residential areas.

Fig.1 Haraz watershed and its location in Iran

2.2. Geospatial database preparation


According to the literature review (e.g, Cao et al., 2016; Chapi et al., 2017; Rahmati and
Pourghasemi, 2017; Tehrany et al., 2015; Zhao et al., 2018) and with regard to the local
characteristics, eleven main factors which have significantly ‎affected the occurrence of flood in
the study area were recognized. Conditioning factors include slope degree, curvature, elevation,
topographic wetness index (TWI), stream power index (SPI), distance to river, river density, land
use, Normalized Difference Vegetation Index (NDVI), rainfall, and lithology. A Digital
Elevation Model (DEM) with 20 m spatial resolution for the study area was obtained from the
Mazandaran Regional Water Authority (MRWA). Since most of the factors were derived from
DEM layer, spatial resolution of all factors was set identical to the DEM layer. On the other
hand, the selected cell size maintains details and produces high precision outputs. (e.g., Chapi et
al., 2017; Hong et al., 2018; Tehrany et al., 2015). Based on the DEM, slope degree, curvature,
elevation, TWI, SPI, distance to river, and river density maps were extracted using ArcGIS 10
(Fig. 2).

Slope degree is related to the water infiltration and the higher velocity so that the higher velocity
of water and the lower rate of water infiltration is associated with the higher slope angle
(Khosravi et al., 2016). Therefore, the lower areas are more exposed to floods. In the present
study, slope angle varies from 0˚ - 66.8˚. Curvature is considered in most literatures as one of the
conditioning factors influencing flooding in basins (Tehrany et al., 2013). A positive value
represents convex surface whereas a negative value indicates ‎concave surface at a given pixel‎
(Chapi et al., 2017). In our study area, curvature map varied from ‎-7.93 - 9.83‎.‎ Elevation is
defined as the height above the sea level, which has relationship with climatic conditions. Hence,
it is an important conditioning factor in flooding of a basin (Tehrany et al., 2015). In this study, it
‎varies from 328 - 5595‎ m.‎ Another factor is TWI, one of the most effective factors on flood
occurrence ‎(Chapi et al., 2017). The ratio between specific basin area and slope is called TWI
(Gallant, 2000), which was calculated as (Beven and Kirkby, 1979):

  
TWI  ln   (1)
 tan  

where a is the cumulative upslope area of a drainage basin through a point (per unit contour
length) and tan β is the angle of slope at the same point. The TWI ranged between 1.9 and 11.‎5.
The SPI (Tehrany et al., 2014), another important factor, was extracted from DEM and was
computed as (Moore and Wilson, 1992):
SPI  As tan 
(2)

is the specific basin area, and  is the local slope gradient (in degree). The SPI range
AS
where
for this study area was from 0-6 to 900,000.
Distance to river as another flood conditioning factor is directly related to closeness and
proximity to main stream in the catchment area (Tehrany et al., 2015). In this study, the river
network was extracted from topographical maps of the Haraz watershed and then divided into
several river types whose length ranged from 0 to 4,352 m. Another effective factor was river
density that is defined as the total length of streams and rivers divided by the total area of the
‎drainage basin indicating how well or poorly a watershed is drained (Oikonomidis et al., 2015).
‎The river density values of the study area ranged from 0 to 2.08 m/m2. Land use is considered as
one of the most important factors on flooding of a basin, which affects the infiltration and runoff
through the nature of surface materials (Kia et al., 2012). ‎

The land use map of the study area was extracted from Landsat-8 Operational Land Imager
(Landsat-8 OLI) satellite images using a neural network algorithm in ENVI 5.1 software (Exelis
Visual Information Solutions, Boulder, Colorado). The obtained map was classified into seven
categories including water bodies, forest, barren, residential areas, woodland, range land and
crop land. The NDVI, another factor, is a valuable index to assess vegetation cover and their
effects on flooding in a basin. This index generally ranges between -1 and +1. The NDVI values
‎ranged from ‎-0.68 to 0.72‎‎. Rainfall was also considered as an important hydrologic process for
recharging basins and flooding in aquifers (Knebl et al., 2005). As rainfall increases, the
potential of flood occurrence increases correspondingly (El Alfy, 2016; Wang et al., 2013). The
rainfall map of the present study was constructed ‎by ordinary kriging interpolation method using
the mean annual rainfall data of 17 rainfall gauges in a period of 20 years (1991–2011). ‎The
rainfall map of the study area ranged between 187.7 mm and 740.5 mm. Lithology, as another
important flood conditioning factor, is related to both soil porosity and water permeability of
aquifers (Tehrany et al., 2014), which in turn influences the flooding in basins in the light of their
lithology. The lithology map of this study was extracted from the geological map at 1:100,000
‎scale obtained from the Geological Survey & Mineral Exploitation of Iran (GSI). ‎The lithology
map of the present study has six units ‎including Jurassic, Teryas, Permian, Quaternary, Tertiary,
Cretaceous (Table 1)

Fig. 2 Flood conditioning factor; a) Curvature, b) Elevation, c) Land cover, d) Lithology, e) NDVI, f)
Rainfall, g) River density, h) Distance to river, i) Slope, j) SPI and k) TWI

Table 1 Data layers and their ranges applied in the flood modeling in this study
Conditioning
factor Range/class types Source
River density 0 - 2.08 DEM
River distance 0 – 4,352 (m) DEM
Slope 0 - 66.8 (%) DEM
TWI 1.9 - 11.5 DEM
Water bodies, Barren land, Forest, Residential areas,
Land cover woodland, range land, crop land Landsat 8
Geological survey
Lithology Jurassic, Teryas, Permian, Quaternary, Tertiary, Cretaceous of Iran
Plan curvature -7.93 - 9.83 DEM
Elevation 328-5,595 DEM
NDVI -0.68 - 0.72 Landsat 8
Rain gauges
Rainfall 187.7 - 740.5 (mm) statistics
SPI 0-6,900,000 DEM

2.3. Individual models


2.3.1. Artificial neural networks (ANN)

The ANN consists of an interconnected set of elements called neuron and a number of layers
(Rahmati and Pourghasemi, 2017; Zhang and Goh, 2016) including the input, hidden and output
layers. In this study, 10 nodes, identical to the number of explanatory variables, were set for the
input and hidden layers, while we used one node for the output layer coded with either 1 for
places where flood occurrence was recorded or 0 for non-occurrence flood. Different algorithms
exist to train (determine the weights) of an ANN model, with Back Propagation (BP) being one
of the most popular ones. Therefore, a BP based ANN was used to approximate the non-linear
relationships between the flood occurrence and exploratory variables. (Zhang and Goh, 2016).
BP selects the initial weights randomly, then compares the calculated values with observed
values and their difference is summarized and reported as mean squared error (Pijanowski et al.,
2002). Based on a generalized delta rule (Rumelhart et al., 1986), the initial weights are then
adjusted so that the total error is distributed among the various neurons in the network
(Pijanowski et al., 2002). This process is iteratively repeated until the error levels off at a low
level.

2.3.2. Classification and regression trees (CART)


The CART is a tree-based model used for both classification and regression tasks (Breiman et
al., 1984). A nonlinear technique that operates based on a set of binary decision rules was
extracted from a number of randomly selected predictors (Breiman et al., 1984). It recursively
partitions the exploratory variables to find the ranges that minimize the residual sum of the
square or misclassification rate. The task of segmentation is controlled by the Gini index, an
index used for calculating the impurity of a data partition by measuring misclassification rates or
sum of the squares errors (Araújo et al., 2011). It has a straightforward and easy to follow
structure so that its generated rules can be easily understood.

2.3.3. Flexible Discriminant Analysis (FDA)

The FDA is a method to accomplish Linear Discriminant Analysis (LDA) on derived responses
using linear regression (Friedman et al., 2001). LDA tries to find a projection hyperplane aiming
at minimizing the interclass variance and maximizing the distance between the projected means
of the classes (Xanthopoulos et al., 2013). The problem of optimization of LDA is to minimize
the square average of the residuals, while, in the FDA, linear regression is replaced by
nonparametric regression. This process will lead to a nonparametric and flexible alternative to
LDA (Friedman et al., 2001). FDA is also a better classifier than LDA for several reasons
(Hastie et al., 2009). Main reason is that LDA has a linear decision boundary that do not
appropriately separate the classes (flood vs. non-flood). Many natural phenomena like flood
(susceptibility) could be non-linear. FDA solves this issue (and other mentioned in the reference)
by generalising LDA to be more flexible. Thus, we used FDA instead of LDA for complex and
nonlinear relations of flood susceptibility mapping.

2.3.4. Generalized Linear Model (GLM)

The GLM with a logistic link function (logistic regression) has been widely used for modelling
binary variables or presence-absence data. In this paper, a logistic regression (LR) to associate
the flood occurrence with underlying environmental driving forces was used. Apart from
creating a susceptibility map that shows the relative likelihood of a cell to be assigned as
torrential, the magnitude of driving forces can also be obtained. The point here is that
interpretation of the obtained coefficient from GLM is not similar to the linear models because
the relationship between variables is not a straight line (James et al., 2013). The model can well
handle both the categorical and continuous variables. Further, response curve for each factor of
the flood can be created.

2.3.5. Generalized Additive Model (GAM)

In flood susceptibility mapping, the relationship between flood and environmental factors is
often complex. GAM is a nonparametric extension of GLM (using kernel or spline smoothers)
and the main difference between them is the ability of GAM to model nonlinear relationships,
and it does not require normal distribution of the dependent (response variable) and the
prediction is generated from a linear combination of predictor variables connected to the
dependent variable using a link function (Hastie and Tibshirani, 1990). Each factor in a GAM
can take the linear or nonlinear forms (Goetz et al., 2011). GAM models have become popular
due to their predictive power, flexibility and interpretability (Zhang and Batterman, 2010).

2.3.6. Boosted Regression Trees (BRT)

The BRT is an ensemble method for fitting statistical models that benefit from combination of
two algorithms: regression trees and boosting (Elith et al., 2008). Handling exploratory variables
of different types, treatment of missing data, handling outliers and insensitive to data
distribution, dealing with complex nonlinear relations and accounting for interaction among the
exploratory variables are key characteristics of a BRT model (Elith et al., 2008). The BRT can be
resembled to the random forest algorithm where both algorithms are constructed from a large
number of trees and overcome limitations of a single tree model. There are two parameters that
need to be set: the „learning rate‟ to determine the involvement of each tree to the growing model
and „tree complexity‟ to control if the interactions are fitted (Buston and Elith, 2011). These
parameters were optimally tuned using the process of cross validation. Furthermore, the effect of
multiple predictors on flood occurrence was investigated using BRTs. In doing so, the relative
influence of each predictor is assessed based on the frequency of the predictor selection and the
model improvement selection.

2.3.7. Multivariate adaptive regression splines (MARS)


Introduced by Friedman (1991), the MARS is a flexible regression method which is used to
model nonlinear relations. It shows a high potential in handling complex and high-dimensional
data (Friedman, 1991). It is a nonparametric model composed of a series of piecewise linear
segments (splines) of differing gradients (Zhang and Goh, 2013). The required basic functions
and their related parameters are automatically determined by the data (Friedman, 1991). MARS
is the product of a two-step process so that the forward step creates functions and detects
potential knots and the backward step comprises pruning. These two steps not only strike a
balance between the number of functions and knots, but also prevents over fitting. The end
points of the segments are called knots or values of a variable that define variation point along
the range of a predictor (Leathwick et al., 2006). MARS generates basic functions by searching
in a stepwise manner. Locating the knot points is performed using adaptive regression algorithm
(Zhang and Goh, 2013). It works under no specific assumption about the data distribution or
relations, a good characteristic for complex spatial data such as the flood.

2.3.8. Maximum entropy (MaxEnt)

The MaxEnt is a method used for species distributions modelling (SDM) from presence-only
species records (Elith et al., 2011). Available since 2004, MaxEnt has been widely employed for
modelling species distributions and its predictive performance is usually competitive with the
methods of highest performing results (Elith et al., 2006). The SDM estimates the relationship
between the influences of environmental and/or spatial factors on species records at specific sites
(Franklin, 2010). The same approach was adopted to assess the influence of environmental
factors on flood occurrence in this study. A total number of 10,000 background points from the
raster layers to represent the available landscape for flooding to be used as presence-background
modelling were randomly selected.

2.4. Ensemble forecasting


Many studies have been working relying on single statistical or ML techniques, however, there is
no single absolute best modeling technique for spatial prediction. In addition, one major source
of uncertainty in spatial prediction is the choice of modelling technique as each model responses
to geographic range and sample size differently (Meller et al., 2014). Nevertheless, combination
of multiple techniques shows that they create robust and more stable results than single
techniques which is called ensemble forecasting or modelling (Araújo and New, 2007; Burnham
and Anderson, 2003). Ensemble modelling has been successfully applied in the spatial prediction
of plant and animal species distributions that outperforms single modelling techniques and
provide a more reliable prediction (Araújo et al., 2005; Breiner et al., 2017; Marmion et al.,
2009). In this paper, ensemble forecasting creates an output from the results of multiple models
so that the errors risen by each model are even and balanced. Seven different ensemble
forecasting approaches available within the R programming package of 'biomod2' (Thuiller et al.,
2009) were implemented including:

i) EMca, Ensemble Model committee averaging which linearly combines the model outputs and
is identical to arithmetic averaging;

ii) EMciInf, Ensemble Model confidence interval Inferior values. In this model, inferior values of
confidence interval maps of each individual model are taken;

iii) EMciSup, Ensemble Model confidence interval Superior values. In contrast to EMciInf, here,
superior values of confidence interval maps of each individual model are taken;

iv) EMcv, Ensemble Model to estimate the coefficient of variation across predictions;

v) EMmean, Ensemble Model to estimate the mean probabilities across predictions;

vi) EMmedian, Ensemble Model median probabilities across predictions;

vii) EMwmean, Ensemble Model weighted mean probabilities across predictions. For this
approach, weight is indeed the Area Under the Receiver Operating Characteristic curve
(AUROC) value obtained for each model multiplied by their corresponding probability map.

Whole the modelling processes were coded in R programming language (R Development Core
Team, 2017).

3. Performance evaluation

When a model is implemented to estimate the spatial likelihood of flood in a given region, its
performance should be tested and validated. In this study, the Relative Operating Characteristics
(ROC) curve was used to assess the performance of the individual and ensemble models. With
caution, it can be stated that in most geo-hazard modelling studies, the ROC has been used as a
standard tool (Pham et al., 2016b; Shahabi and Hashim, 2015). The ROC curve is designed in a
two-dimensional space in which the X-axis specifies sensitivity and the Y-axis denotes
specificity. The sensitivity is defined as the number of flood pixels correctly classified as flood,
while the specificity is the number of non-flood pixels correctly classified as non-flood (Bui et
al., 2016). As an integral part of ROC curve, the Area Under the Receiver Operating
Characteristic (AUROC) is a quantitative metric commonly used to evaluate the performance of
the flood models (Pham et al., 2016a). When the value of AUROC is 0.5, it shows that the model
operates like a random model and the values greater than 0.5 up to 1 indicate the higher
performance of a given model (Shirzadi et al., 2017). For further description of ROC, one can
refer to Pontius and Parmentier (2014).

4. Results and discussion


4.1 Influence of spatial factors on the relative likelihood of flood occurrence
The relationship among 11 conditioning factors and flood occurrence was assessed using the
BRT model. Figure 3 shows this relationship in which the X-axis represents the distribution of
pixel values of each conditioning factor and the Y-axis denotes the relative likelihood of flood
occurrence. These results clearly imply that distance from river can be considered as the most
effective conditioning factor on FMS, followed by NDVI, elevation, rainfall, slope, land cover,
SPI, river density, TWI, lithology and curvature. This is partly in agreement with findings
reported by Chapi et al. (2017), even though slope was reported as the main effective factor.

In the curvature factor (Fig 4.a), the probability of flood occurrence increases between -1 and 0
values. This range of curvature conforms to the flat zones. The result indicates that curvature
values less than -1 which represents convex slopes (negative value of curvature), have lower
probability of flood occurrence (~ 0.4) with a constant trend than the concave slopes that flood
probability increases to 0.6 in those regions (curvature values more than 0 or positive value of
curvature). General response of curvature values to flood occurrence confirms the results
reported by Cao et al., (2016), Chapi et al., (2017) and Khosravi et al. (2016).
As the elevation (Fig 4.b) increases, the probability of flooding decreases. Relationship between
elevation and the relative likelihood of flood occurrence is negative up to 2700 m, meaning the
chance of flood occurrence is higher in lower lands than elevations higher than 2700 m. The
most probability of flood occurrence (0.8) is shown to occur in elevations around 300 m. Our
result is in line with the findings reported by (Hong et al., 2017).
The role of land cover (Fig 4.c) in flood occurrence in the study area was also quite remarkable.
Amongst the land cover classes, water bodies and barren lands displayed the most and least
impacts on flooding and showed the probability of flood occurrence close to 0.7 and 0.5,
respectively. Naturally, barren lands are more vulnerable to flooding, because there is no
vegetation to slow down the water flow and increase the penetration rate. On the other hand, as
expected, the forest showed the lowest impact on the flood among all land cover classes.
In terms of lithology (Fig 4.d), the areas with high resistant rocks or highly permeable subsoil
materials have lower drainage density as well as lower probability of flood occurrence
(Srivastava et al., 2014). In the study area, although tertiary formations have higher value of the
probability of flood occurrence (0.6), other lithology formations showed almost the same
importance for flooding (~ 0.6).
NDVI (Fig 4.e) reflects the density of vegetation cover, although depending on the purpose of
the study, various interpretations could be inferred. NDVI affects the flood in different ways
under different conditions. Higher NDVI and dense vegetation reduce and slow water flow
(Turoğlu and Dölek, 2011). The dense vegetation gives the water time to penetrate into the
ground, resulting in a decrease in water volume and less probability of flood event (Turoğlu and
Dölek, 2011). In this study, the effect of NDVI on flood incidence is very considerable. With a
glimpse into Figure 3, it can be found that the lower the NDVI value, the higher the probability
of flood occurrence will be. Additionally, higher probability of flood incidence (~ 0.9) in our
study area was obtained by the NDVI less than -0.5. In a nutshell, the response curve shows that
less covered areas are more at the risk of flood.
It has been confirmed in the previous studies that the higher the rainfall (Fig 4.f), the higher the
probability of flood incidence will be (Todini et al., 2004). In this study, the rainfalls up to 400
mm had the most important role in flooding, while rainfalls over 400 mm showed a constant
effect. The rainfalls approximately around 220 mm are associated with the highest chance of
flood occurrence (0.8) in our study area.
Another factor which can directly affect the probability of flood occurrence is river density (Fig
4.g) (Chapi et al., 2017). In this study, river densities between 0.8 and 2 m/km2 were more
causative in flooding. The higher the river density (>0.8), the higher the probability of flood
incidence will be. Also, river density approximately around 1.1 m/km2 played the greatest role in
flood occurrence in the Haraz watershed with the highest probability (~ 0.61)
Since most of floods occur due to overflowing of water from river banks, distance to river (Fig
4.h) is a critical factor for spatial prediction of flood prone areas in a watershed (Chapi et al.,
2017). Therefore, the areas closer to rivers demonstrate rapid response to rainstorms and,
therefore, are more exposed to flooding (Butler et al., 2006). In this study, the results indicated
that distance to rivers up to 150 m was the most influential distance to flood occurrence. The
lower the distance to the river (~ <20 m), the higher the probability of flood occurrence (~ 0.9)
will be. Our findings confirm the results obtained by Hong et al. (2017), where the class of 0-100
m from the river showed a higher association in relation to flood occurrence.
The Slope (Fig 4.i) response curve shows that the lower the slope angle, the higher the
probability of flood occurrence will be, a finding that is consistent with that of other studies (e.g.,
Rahmati and Pourghasemi, 2017; Tehrany et al., 2014). It implies that the chance of flood
occurrence increases in flat areas. Figure 4 shows that the most probable slope range for the
flood occurrence falls between 0- 20˚, so that the probability of flood occurrence decreases after
20˚ and it drops and levels off for the slopes steeper than about 20˚. In particular, the likelihood
of flood occurrence at falt areas and slopes up to 8˚ is extremely high while it falls sharply at
slopes of more than 8˚.
The SPI and TWI as two effective hydrological factors can influence the spatial variation of
floods. SPI (Fig 4.j) represents soil water content and erosion power of floods to flow
downwards in a watershed (Cao et al., 2016). The lower SPI is associated with the stronger
power of flood occurrence. The areas with ability of flow accumulation are specified with the
lower values of SPI (Turoğlu and Dölek, 2011). In this regard, the SPI close to zero value has
higher probability of flood occurrence (> 0.5) indicating the most floods in the study area have
occurred in places with lower values of SPI. (Beven, 2011) have reported that the TWI (Fig 4.k)
implies saturated condition as well as water accumulation in a watershed. Therefore, the areas of
a watershed identified by higher values of TWI are more exposed to flood inundation. In this
study, the TWI showed a fixed role with almost no spatial variability in the occurrence of the
flood.
Fig. 3 The relationship between conditioning factors and the probability of flood occurrence.
Land cover codes correspond to 1: water bodies, 2: barren land, 3: forest, 4: residential areas, 5:
wood land, 6: range land, 7: crop land. Lithology codes correspond to 1: Jurassic, 2: Teryas, 3:
Permian, 4: Quaternary, 5: Tertiary, 6: Cretaceous

4.2. Flood susceptibility mapping using individual and ensemble models

In the current study, all individual models were first implemented and then, based on the results
of individual models, the ensemble forecasting approaches were derived. The result of both
model categories was a flood susceptibility map. Figures 5 and 6 show the flood susceptibility
maps generated by the individual and ensemble models. According to the response curve
obtained by the BRT model, there is a chance of flood occurrence only in the vicinity of 500 m
of the rivers. Therefore, in order to model the likelihood of a flood event, a 500 m buffer zone
was considered. With regard to the scale of the map, it is difficult to visualize the details in a 500
m buffer zone when displaying the entire map. As a result, an area in which the density of rivers
was considerable was selected and magnified (Figures 4 and 5). As seen in Figures 4 and 5, the
output of all prediction maps vary between 0 and 1. The values near to zero and 1 indicate the
lowest and highest likelihood of flood occurrence respectively. As the figures show, individual
models show roughly different patterns as compared to ensemble ones, a problem that the
Standard Deviation (S.D) map also confirms. The map of standard deviation was created by
running the standard deviation among the individual models and showed a remarkable variability
among the prediction of individual models, indicating that the model operates well in some and
in some places it produces a high error. This indicates the sensitivity of individual models while
the ensemble models show a more uniform pattern excluding the EMcv.
Fig. 4 Flood susceptibility maps extracted from individual models for the Haraz watershed

Fig. 5 Flood susceptibility maps extracted from ensembles models in the Haraz watershed

4.3 Performance of individual and ensemble models

In this study, eight individual models (ANN, CART, FDA, GLM, GAM, BRT, MARS and
MaxEnt), and seven ensemble models (EMca, EMciInf, EMciSup, EMcv, EMmean, EMmedian
and EMwmean) were calibrated and implemented for flood susceptibility mapping. The
AUROC, S.D and significant level (Sig.) were utilized for assessing the performance of these
models. Results can be observed in Tables 2 and 3. The values of AUROC in both individual and
ensemble models were statistically significant due to having Sig. equal to 0.000 (less than 0.05).
For the ease of interpretation, performance of individual and ensemble models was overlaid on a
single ROC (Figure 6a and b). Accordingly, the best model performance among the individual
models belonged to the BRT (AUROC=0.975). It was followed by the GLM and MaxEnt
(AUROC=0.971), the GAM model (AUROC=0.962), the MARS model (AUROC=0.941), the
ANN model (AUROC=0.920), the FDA model (AUROC=0.822), and the CART model
(AUROC=0.643). Results also indicated that the EMmedian ensemble model (AUROC=0.975)
among the ensemble models had the highest performance, followed by the EMmean and
EMwmean models (AUROC=0.973), the EMciSup model (AUROC=0.972), the EMciInf model
(AUROC=0.869), and the EMcv model (AUROC=0.732).

Table. 2 AUROC, standard deviation (S.D) and significant level (Sig.) of individual
models
Variables AUC S.D Sig. Asymptotic 95% Confidence Interval
Lower Bound Upper Bound
ANN 0.920 0.027 0.000 0.867 0.973
CART 0.643 0.047 0.000 0.551 0.735
FDA 0.822 0.040 0.000 0.743 0.901
GLM 0.971 0.005 0.000 0.960 0.981
GAM 0.962 0.014 0.000 0.934 0.990
BRT 0.975 0.007 0.000 0.961 0.989
MARS 0.941 0.016 0.000 0.909 0.973
MaxEnt 0.971 0.006 0.000 0.959 0.982

Table 3 AUROC, standard deviation (S.D) and significant level (Sig.) of ensemble

Variables AUC S.D Sig. Asymptotic 95% Confidence Interval


Lower Bound Upper Bound
EMca 0.959 0.013 0.000 0.933 0.985
EMcInf 0.870 0.035 0.000 0.802 0.938
EMciSup 0.972 0.006 0.000 0.961 0.983
EMcv 0.733 0.042 0.000 0.650 0.816
EMmean 0.974 0.005 0.000 0.963 0.984
EMmedian 0.976 0.005 0.000 0.966 0.986
EMwmean 0.974 0.005 0.000 0.963 0.984
Fig. 6 Comparison specify the prioritization of performance of individual and ensemble
models

4.4 Comparison of individual and ensemble forecasting


The prediction power among the best individual and ensemble models was graphically
determined using the ROC curve (Fig. 7). Basically, the highest value of the area under the ROC
curve among individual models was obtained by the BRT (AUC=0.975), while the EMmedian
model acquired the highest value of AUC (0.976) among the ensemble models. As shown, there
is no outstanding difference between the two models.
BRTs can be considered as an advanced form of regression although they belong to ML
techniques (Friedman et al., 2000; Schapire, 2003). BRT is composed of regression trees and
boosting and, in contrast to the regression models that output a single predictive model, they fit
numerous simple models and combine them for prediction (Buston and Elith, 2011). BRT can
successfully consider different types of variables including numeric, binary, categorical, and
missing values as well as non-independent data. Further, it automatically models the interactions
between predictors (Elith et al., 2008). In contrast to the other individual models used in this
research, the BRT operation is based on a boosting approach where the aim is improving the
final prediction and accuracy. Boosting generates a more robust final model and allows
curvilinear functions to be modelled (Elith et al., 2008).
Fig. 7 Comparison of the prediction power between the best individual and ensemble models for
flood susceptibility mapping in the Haraz watershed

5. Lessons learned from the application of six flood susceptibility models

Flood susceptibility mapping can be implemented using a verity of models. The spatial pattern of
predicted map in different modelling techniques might be different, although their evaluation
might be similar while using the AUROC value. Therefore, selecting one model as “the best” for
creating flood susceptibility mapping and decision making is quite challenging. Model averaging
or ensemble forecasting is a technique to combine the results of different individual models in a
more robust single model (Araújo and New, 2007). This approach is especially powerful when
applying the calibrated model to new regions since it maps the main trend among individual
modelling techniques (e.g. median and mean) and catches overall variations, hence reducing
uncertainty in the spatial prediction (Guisan et al., 2017). There are several ways to combine the
individual models (Marmion et al., 2009); however, there is no guarantee that ensemble models
always outperform individual models or boost the AUC values of the individual models
impressively. AUC value should be interpreted as the power of a model in realization of the
occurrence versus non-occurrence. Important aspects to consider are generalization ability and
creation of a model more stable and less prone to over fitting, the matters that are substantially
covered by ensemble forecasting. Furthermore, the best individual model does not assure that, by
changing the input data or changing the future conditions, it can still generate the highest
accuracy.

6. Conclusion
To alleviate the devastating effects of floods, creating accurate flood susceptibility maps is
essential. In this study, eight individual ML and statistical techniques for mapping flood
susceptibility were implemented and cross compared. The integration of individual models has
then resulted in the emergence of seven ensemble forecasting approaches, leading to the
application of 15 different approaches for mapping flood susceptibility. Our uncertainty map,
which resulted from running the standard deviation among the individual models, showed an
outstanding variability among the prediction of individual models, indicating that each model
predicts well for some places and generates less accurate predictions for the other areas. Model
evaluation showed that the BRT model had the highest prediction power among individual
models, whereas the EMmedian appeared to be the most accurate model among the ensemble
ones, although the difference between the value of AUROC for the BRT and EMmedian models
was not significant. On the other hand, investigation of spatial influence of each factor on the
flood occurrence showed the distance to river, NDVI, elevation, and rainfall as the most effective
factors. As a conclusion, ensemble models are suggested for flood mapping due to their more
stable results, more generalization ability, and higher prediction accuracy. Nevertheless, there is
limited literature on the application of ensemble forecasting approaches in flood mapping, and
thus development of other ensemble frameworks such as model stacking is strongly
recommended.

Acknowledgment
Authors wish to acknowledge the financial support of the Iran National Science Foundation
(INSF) through the research project No 96004000. Roozbeh Valavi is supported by an Australian
Government Research Training Program Scholarship and a Rowden White Scholarship. Also,
the authors would like to acknowledge the anonymous reviewers and editor for their ‎helpful
comments on a previous version of the manuscript.‎

References
Araújo, M., Rivas, T., Giraldez, E., Taboada, J., 2011. Use of machine learning techniques to analyse the
risk associated with mine sludge deposits. Mathematical and Computer Modelling 54, 1823-1828.
Araújo, M.B., New, M., 2007. Ensemble forecasting of species distributions. Trends in ecology &
evolution 22, 42-47.
Araújo, M.B., Whittaker, R.J., Ladle, R.J., Erhard, M., 2005. Reducing uncertainty in projections of
extinction risk from climate change. Global ecology and Biogeography. 14, 529-538.
Beven, K., Kirkby, M.J., 1979. A physically based, variable contributing area model of basin
hydrology/Un modèle à base physique de zone d'appel variable de l'hydrologie du bassin versant.
Hydrological Sciences Journal 24, 43-69.
Beven, K.J., 2011. Rainfall-runoff modelling: the primer. John Wiley & Sons.
Breiman, L., Friedman, J., Stone, C.J., Olshen, R.A., 1984. Classification and regression trees. CRC
press, New York‎.
Breiner, F.T., Nobis, M.P., Bergamini, A., Guisan, A., 2017. Optimizing ensembles of small models for
predicting the distribution of species with few occurrences. Methods in Ecology and Evolution.
Bui, D.T., Tuan, T.A., Klempe, H., Pradhan, B., Revhaug, I., 2016. Spatial prediction models for shallow
landslide hazards: a comparative assessment of the efficacy of support vector machines, artificial
neural networks, kernel logistic regression, and logistic model tree. Landslides 13, 361-378.
Burnham, K.P., Anderson, D.R., 2003. Model selection and multimodel inference: a practical
information-theoretic approach, 2nd ed. Springer Science & Business Media.
Buston, P.M., Elith, J., 2011. Determinants of reproductive success in dominant pairs of clownfish: a
boosted regression tree analysis. Journal of Animal Ecology 80, 528-538.
Butler, D., Kokkalidou, A., MAKROPOULOS, C., 2006. Supporting the siting of new urban
developments for integrated urban water resource management. Integrated Urban Water Resources
Management, 19-34.
Cao, C., Xu, P., Wang, Y., Chen, J., Zheng, L., Niu, C., 2016. Flash flood hazard susceptibility mapping
using frequency ratio and statistical index methods in coalmine subsidence areas. Sustainability 8, 948.
Chapi, K., Singh, V.P., Shirzadi, A., Shahabi, H., Bui, D.T., Pham, B.T., Khosravi, K., 2017. A novel
hybrid artificial intelligence approach for flood susceptibility assessment. Environmental Modelling &
Software 95, 229-245.
El Alfy, M., 2016. Assessing the impact of arid area urbanization on flash floods using GIS, remote
sensing, and HEC-HMS rainfall–runoff modeling. Hydrology Research 47, 1142-1160.
Elith, J., Graham, C.H., Anderson, R.P., Dudík, M., Ferrier, S., Guisan, A., J Hijmans, R., Huettmann, F.,
R Leathwick, J., Lehmann, A., Li, J., G Lohmann, L., Overton, J.M.M., Peterson, A.T., Phillips, S.J.,
Richardson, K., Scachetti-Pereira, R., Schapire, R.E., Soberón, J., Williams, S., Wisz, M.S.,
Zimmermann, N.E., 2006. Novel methods improve prediction of species' distributions from occurrence
data. Ecography 29, 129-151.
Elith, J., Leathwick, J.R., Hastie, T., 2008. A working guide to boosted regression trees. Journal of
Animal Ecology 77, 802-813.
Elith, J., Phillips, S.J., Hastie, T., Dudík, M., Chee, Y.E., Yates, C.J., 2011. A statistical explanation of
MaxEnt for ecologists. Diversity and distributions 17, 43-57.
Franklin, J., 2010. Mapping species distributions: spatial inference and prediction. Cambridge University
Press.
Friedman, J., Hastie, T., Tibshirani, R., 2000. Additive logistic regression: a statistical view of boosting
(with discussion and a rejoinder by the authors). The annals of statistics 28, 337-407.
Friedman, J., Hastie, T., Tibshirani, R., 2001. The elements of statistical learning. Springer series in
statistics New York.
Friedman, J.H., 1991. Multivariate adaptive regression splines. The annals of statistics, 1-67.
Gallant, J.C., 2000. Terrain analysis: principles and applications. John Wiley & Sons.
Goetz, J.N., Guthrie, R.H., Brenning, A., 2011. Integrating physical and empirical landslide susceptibility
models using generalized additive models. Geomorphology 129, 376-386.
Guisan, A., Thuiller, W., Zimmermann, N.E., 2017. Habitat Suitability and Distribution Models: with
Applications in R. Cambridge University Press.
Hastie, T., Tibshirani, R., 1990. Generalized additive models. Wiley Online Library.
Hastie, T., Tibshirani, R., Friedman, J., 2009. The elements of statistical learning: Data Mining, Inference,
and Prediction, 2nd ed. Springer series in statistics New York.
Hong, H., Panahi, M., Shirzadi, A., Ma, T., Liu, J., Zhu, A.-X., Chen, W., Kougias, I., Kazakis, N., 2017.
Flood susceptibility assessment in Hengfeng area coupling adaptive neuro-fuzzy inference system with
genetic algorithm and differential evolution. Science of The Total Environment.
Hong, H., Tsangaratos, P., Ilia, I., Liu, J., Zhu, A.-X., Chen, W., 2018. Application of fuzzy weight of
evidence and data mining techniques in construction of flood susceptibility map of Poyang County,
China. Science of The Total Environment 625, 575-588.
James, G., Witten, D., Hastie, T., Tibshirani, R., 2013. An introduction to statistical learning. Springer.
Khosravi, K., Nohani, E., Maroufinia, E., Pourghasemi, H.R., 2016. A GIS-based flood susceptibility
assessment and its mapping in Iran: a comparison between frequency ratio and weights-of-evidence
bivariate statistical models with multi-criteria decision-making technique. Natural Hazards 83, 947-
987.
Khosravi, K., Pham, B.T., Chapi, K., Shirzadi, A., Shahabi, H., Revhaug, I., Prakash, I., Bui, D.T., 2018.
A comparative assessment of decision trees algorithms for flash flood susceptibility modeling at Haraz
watershed, northern Iran. Science of the Total Environment 627, 744-755.
Kia, M.B., Pirasteh, S., Pradhan, B., Mahmud, A.R., Sulaiman, W.N.A., Moradi, A., 2012. An artificial
neural network model for flood simulation using GIS: Johor River Basin, Malaysia. Environmental
Earth Sciences 67, 251-264.
Knebl, M., Yang, Z.-L., Hutchison, K., Maidment, D., 2005. Regional scale flood modeling using
NEXRAD rainfall, GIS, and HEC-HMS/RAS: a case study for the San Antonio River Basin Summer
2002 storm event. Journal of Environmental Management 75, 325-336.
Leathwick, J., Elith, J., Hastie, T., 2006. Comparative performance of generalized additive models and
multivariate adaptive regression splines for statistical modelling of species distributions. Ecological
modelling 199, 188-196.
Lee, M.-J., Kang, J.-e., Jeon, S., 2012. Application of frequency ratio model and validation for predictive
flooded area susceptibility mapping using GIS, Geoscience and Remote Sensing Symposium
(IGARSS), 2012 IEEE International. IEEE, pp. 895-898.
Marmion, M., Parviainen, M., Luoto, M., Heikkinen, R.K., Thuiller, W., 2009. Evaluation of consensus
methods in predictive species distribution modelling. Diversity and distributions 15, 59-69.
Meller, L., Cabeza, M., Pironon, S., Barbet-Massin, M., Maiorano, L., Georges, D., Thuiller, W., 2014.
Ensemble distribution models in conservation prioritization: from consensus predictions to consensus
reserve networks. Diversity and distributions. 20, 309-321.
Merkuryeva, G., Merkuryev, Y., Sokolov, B.V., Potryasaev, S., Zelentsov, V.A., Lektauers, A., 2015.
Advanced river flood monitoring, modelling and forecasting. Journal of Computational Science 10,
77-85.
Moore, I.D., Wilson, J.P., 1992. Length-slope factors for the Revised Universal Soil Loss Equation:
Simplified method of estimation. Journal of soil and water conservation 47, 423-428.
Oikonomidis, D., Dimogianni, S., Kazakis, N., Voudouris, K., 2015. A GIS/Remote Sensing-based
methodology for groundwater potentiality assessment in Tirnavos area, Greece. Journal of Hydrology
525, 197-208.
Pham, B.T., Bui, D.T., Prakash, I., Dholakia, M., 2016a. Rotation forest fuzzy rule-based classifier
ensemble for spatial prediction of landslides using GIS. Natural Hazards 83, 97-127.
Pham, B.T., Pradhan, B., Bui, D.T., Prakash, I., Dholakia, M., 2016b. A comparative study of different
machine learning methods for landslide susceptibility assessment: a case study of Uttarakhand area
(India). Environmental Modelling & Software 84, 240-250.
Pijanowski, B.C., Brown, D.G., Shellito, B.A., Manik, G.A., 2002. Using neural networks and GIS to
forecast land use changes: a land transformation model. Computers, environment and urban systems
26, 553-575.
Pontius, R.G., Parmentier, B., 2014. Recommendations for using the relative operating characteristic
(ROC). Landscape ecology 29, 367-382.
Pourghasemi, H.R., Pradhan, B., Gokceoglu, C., 2012. Application of fuzzy logic and analytical hierarchy
process (AHP) to landslide susceptibility mapping at Haraz watershed, Iran. Natural hazards 63, 965-
996.
Rahmati, O., Pourghasemi, H.R., 2017. Identification of Critical Flood Prone Areas in Data-Scarce and
Ungauged Regions: A Comparison of Three Data Mining Models. Water Resources Management 31,
1473-1487.
Rumelhart, D.E., Hinton, G.E., Williams, R.J., 1986. Learning representations by back-propagating
errors. nature 323, 533.
Schapire, R.E., 2003. The boosting approach to machine learning: An overview, Nonlinear estimation and
classification. Springer, pp. 149-171.
Shahabi, H., Hashim, M., 2015. Landslide susceptibility mapping using GIS-based statistical models and
Remote sensing data in tropical environment. Scientific reports 5, 9899.
Shirzadi, A., Bui, D.T., Pham, B.T., Solaimani, K., Chapi, K., Kavian, A., Shahabi, H., Revhaug, I., 2017.
Shallow landslide susceptibility assessment using a novel hybrid intelligence approach. Environmental
Earth Sciences 76, 60.
Srivastava, P.K., Han, D., Rico‐Ramirez, M.A., Islam, T., 2014. Sensitivity and uncertainty analysis of
mesoscale model downscaled hydro‐meteorological variables for discharge prediction. Hydrological
processes 28, 4419-4432.
Tang, Z., Zhang, H., Yi, S., Xiao, Y., 2018. Assessment of flood susceptible areas using spatially explicit,
probabilistic multi-criteria decision analysis. Journal of Hydrology.
Tehrany, M.S., Pradhan, B., Jebur, M.N., 2013. Spatial prediction of flood susceptible areas using rule
based decision tree (DT) and a novel ensemble bivariate and multivariate statistical models in GIS.
Journal of Hydrology 504, 69-79.
Tehrany, M.S., Pradhan, B., Jebur, M.N., 2014. Flood susceptibility mapping using a novel ensemble
weights-of-evidence and support vector machine models in GIS. Journal of hydrology 512, 332-343.
Tehrany, M.S., Pradhan, B., Mansor, S., Ahmad, N., 2015. Flood susceptibility assessment using GIS-
based support vector machine model with different kernel types. Catena 125, 91-101.
Thuiller, W., Georges, D., Engler, R., Breiner, F., 2016. biomod2: Ensemble Platform for Species
Distribution Modeling. R package version 3.3-7. https://CRAN.R-project.org/package=biomod2
Thuiller, W., Lafourcade, B., Engler, R., Araújo, M.B., 2009. BIOMOD-a platform for ensemble
forecasting of species distributions. Ecography 32, 369-373.
Todini, F., De Filippis, T., De Chiara, G., Maracchi, G., Martina, M., Todini, E., 2004. Using a GIS
approach to asses flood hazard at national scale. Proceedings of the European Geosciences Union, 1st
General Assembly: 25–30 April 2004; Nice, France.
Turoğlu, H., Dölek, İ., 2011. Floods and their likely impacts on ecological environment in Bolaman River
basin (Ordu, Turkey). Research Journal of Agricultural Science 43, 167-173.
Wang, L.-C., Behling, H., Lee, T.-Q., Li, H.-C., Huh, C.-A., Shiau, L.-J., Chen, S.-H., Wu, J.-T., 2013.
Increased precipitation during the Little Ice Age in northern Taiwan inferred from diatoms and
geochemistry in a sediment core from a subalpine lake. Journal of paleolimnology 49, 619-631.
Xanthopoulos, P., Pardalos, P.M., Trafalis, T.B., 2013. Linear discriminant analysis, Robust Data Mining.
Springer, pp. 27-33.
Zhang, K., Batterman, S., 2010. Near-road air pollutant concentrations of CO and PM 2.5: a comparison
of MOBILE6. 2/CALINE4 and generalized additive models. Atmospheric Environment 44, 1740-
1748.
Zhang, W., Goh, A.T., 2016. Multivariate adaptive regression splines and neural network models for
prediction of pile drivability. Geoscience Frontiers 7, 45-52.
Zhang, W., Goh, A.T.C., 2013. Multivariate adaptive regression splines for analysis of geotechnical
engineering systems. Computers and Geotechnics 48, 82-95.
Zhao, G., Pang, B., Xu, Z., Yue, J., Tu, T., 2018. Mapping flood susceptibility in mountainous areas on a
national scale in China. Science of The Total Environment 615, 1133-1142.

You might also like